I started in recruitment in 2008. Not the year before, not the year after. I was placing people into financial services roles the same month Lehman Brothers filed for bankruptcy, which is a bit like learning to swim by being pushed off a boat.

Eighteen years later I have watched four separate waves come through and rearrange who works and who doesn’t. Each one arrived with its own vocabulary and its own panic. Each one was described at the time as the end of something. And every time, the thing that actually got destroyed was not the thing everyone was watching.

I want to write down what I’ve seen, because I think we are getting this one wrong in a specific and expensive way.

2008 didn’t take jobs. It took a cohort.

The financial crisis was loud. Headcount froze across every bank I worked with, hiring managers stopped returning calls, and roles that had been open for months simply disappeared from the system. Within about three years most of those roles came back.

What didn’t come back was the class of 2008 and 2009.

This is the part people forget. The research on this is unambiguous — Lisa Kahn’s work at Yale on graduates who entered the labour market during a recession found a wage penalty that persisted for well over a decade after the economy recovered. Not a year of lost earnings. A permanent discount, carried for the rest of a career, by people who had done nothing except graduate at the wrong time.

I watched it happen in my own candidate pool. The people who got in just before the crash were three promotions ahead of equally capable people who got in eighteen months later, and the gap never closed. Timing beat talent, and it kept beating it.

That was my first real lesson in this job: a downturn doesn’t distribute its damage evenly. It picks the people with the least tenure and charges them for the whole thing.

Hold that thought.

Automation took the task, then invented a new job to replace it

The 2010s brought a different mechanism. Offshoring, then industrial automation. This time the jobs genuinely did not come back — the factory floor roles, the shared service centre roles, the back-office processing that moved first to Manila and Bangalore and then to software.

The numbers here are real. Acemoglu and Restrepo’s work at MIT found each additional industrial robot replaced roughly 3.3 US workers and pushed local wages down. That is not a projection. That is what already happened.

But something else also happened, and it’s the part the doom coverage consistently missed: the roles that appeared afterwards did not exist before. I was placing cloud architects by 2015. I was placing data engineers by 2017. Nobody had those titles when I started. The job market didn’t shrink — it moved, and it moved faster than the education system could follow.

I came out of that decade believing something that I now think is only half true: that displacement always produces replacement, and that the task is just to help people cross the gap.

Covid moved the work. It didn’t remove it.

2020 was the loudest and the least structural. Everyone assumed the labour market was about to collapse. What actually happened was that work changed location, and then changed power.

For about eighteen months, candidates had leverage they had never had before. I watched salary expectations reset upward across the whole region. Then the leverage quietly went back to employers and most people pretended it hadn’t.

Covid taught me that a shock can look like a revolution and turn out to be a renegotiation.

This wave is not taking the job. It is taking the first rung.

Now the part that is genuinely new, and the reason I’m writing this rather than another AI think-piece.

Every previous wave took work from the bottom or the middle and left the structure intact. The apprenticeship survived. You still started somewhere, did unglamorous work badly, got corrected, and became competent. The ladder lost rungs but it stayed a ladder.

This one is removing the bottom rung specifically.

One of Singapore’s largest banks has already said publicly that there will be no more entry-level roles in the near future. When an institution of that size says it out loud, the rest aren’t disagreeing — they’re just not announcing it yet. I’ve sat in enough client conversations this year to know that the thinking is already there.

And here is what nobody is pricing in. Entry-level work was never really about the output. The output was mediocre, everyone knew it, and we paid for it anyway. What we were actually buying was the production of a senior person. You cannot become a competent operator without three years of doing the work badly under someone who corrects you. Remove that and you don’t lose one cohort of juniors. You lose the supply of every mid-level and senior hire for the next two decades.

I fill those roles. I can tell you with some confidence that the pipeline is a physical thing, it has a lead time of about seven to ten years, and we are currently cutting the input while expecting the output to continue.

Ask a CFO to justify a junior headcount against a software licence and the junior loses every time. The licence is cheaper, faster, and available immediately. The cost of not having a thirty-four-year-old who knows what they’re doing in 2034 does not appear on any budget line in 2026. So it gets ignored, by everyone, simultaneously.

The Philippines is the clearest early warning in our region. The BPO sector employs around 1.9 million people and generates close to 8% of national GDP. Revenue there is still forecast to grow in 2026 — around 5.3% — while employment growth slows to roughly 3.6%. That gap is the entire argument in one line. The industry is not dying. It is becoming an industry that no longer needs people, in a country that built an economy on it needing people.

Gen Alpha will be fluent. That is not the same as capable.

The hopeful version of this story is that Gen Alpha, who will have had AI since infancy, will simply be good at it in a way the rest of us can’t be.

I’m sceptical, and I think the scepticism is the useful contribution here.

We already ran this experiment. Millennials and Gen Z were called digital natives because they grew up with the internet. Fluency with an interface turned out to be almost entirely uncorrelated with the ability to build anything, evaluate anything, or think rigorously about anything. Knowing how to use TikTok did not make anyone a media strategist.

Gen Alpha will be extremely fluent with these tools. They will prompt faster than I ever will and they will not need a manual. But the scarce skill in an AI economy isn’t operating the tool — the tool is designed to need no skill. The scarce skill is judgement: knowing what to ask, recognising when the answer is confidently wrong, and deciding what should be built at all. Those are learned by being wrong in front of someone more experienced, repeatedly, over years.

Which is exactly the apprenticeship we are in the middle of removing.

The countries taking this seriously are acting on it at the policy level. China has begun embedding AI across its entire education system and removing courses it has decided are obsolete. You can argue with the method — it’s top-down and it will get things wrong — but in five years it produces a population trained for the demand that’s coming rather than the demand that’s gone. Most of the region is not doing anything comparable. Education is, by some distance, the industry that has been least changed by technology: we still put children in a room for eight hours and lecture them, which is structurally the same model we had in the industrial age.

And if nobody is earning, who exactly is buying?

This is the question I keep coming back to and have not heard a good answer to.

Businesses exist because people buy things. People buy things because they have income. If AI removes enough income from enough people, the demand side of the economy goes with it — and the companies cutting headcount to improve margins are, in aggregate, cutting their own customers.

The only serious proposal on the table is universal basic income, which is a policy answer to a problem no government has yet admitted it has, and which would require a level of coordination nobody is demonstrating.

I don’t have the answer. I’m not sure anyone does. But I notice that the people selling the technology are not being asked the question, and the people who should be asking it are mostly the ones buying.

What I actually do with this

I’m not going to end with advice about upskilling, because I think most of that advice is a way of making a structural problem sound like a personal failing.

What I’d say instead:

If you’re early career: your advantage is not that you’re fast with the tools. Everyone will be. Your advantage is anything that requires being wrong in front of someone and learning from it. Go where that is still possible, even if the title is worse and the salary is lower, because the alternative is being fluent and uncoachable at thirty.

If you’re hiring: your junior headcount is not a cost line, it’s your 2034 senior bench, and you are making a ten-year decision in a quarterly meeting. At minimum, know that’s what you’re doing.

If you’re mid-career: the thing that protects you is the judgement you accumulated, not the tasks you perform. Most people have this backwards and are defending the tasks.

And I’ll be honest about my own position, because it would be dishonest not to. I run a recruitment business. I have watched AI do parts of my job. I don’t think the work disappears, but I’m certain it changes (or has changed), and anyone in my industry telling you otherwise is selling something.

Four waves in eighteen years. The first three took jobs and gave different ones back.

I’m genuinely not sure this one does. I’d like to be wrong about that, and I’d rather say so now than pretend I saw it coming later.